Biologically Inspired Predictive Coding TCN-Transformer for Anticipatory Human-Robot Interaction in Shared Physical Spaces

TL;DR

This study develops a biologically inspired neural network model to decode EEG signals for predicting human motion intentions, enabling proactive human-robot interaction in shared spaces.

Contribution

It introduces a novel TCN-Transformer architecture inspired by biological predictive coding to decode EEG signals for anticipatory human motion prediction.

Findings

High-beta oscillations in right fronto-central EEG predict motor readiness.

The model achieved an AUC of 0.727 with 1-second look-ahead.

Hemisphere asymmetries in neural activity were identified.

Abstract

As mobile robots increasingly operate in environments shared with humans, proactively anticipating human motion rather than responding reactively is critical for preempting collisions during close-proximity navigation, while maintaining mobility efficiency and avoiding unnecessary yields. A timely and motivating engineering application is how autonomous vehicles interpret ambiguous right-of-way such as unsignalized pedestrian crossings. To address this challenge, this study explores the feasibility of decoding preparatory neural activity from wearable electroencephalography (EEG) to predict human motion intention before it is behaviorally expressed. Drawing inspiration from biological predictive coding mechanisms between the sensorimotor cortex and insula-frontoparietal network, we implement this principle in a Temporal Convolutional Network-Transformer architecture to decode fast-evolving EEG signals underlying perception-action transitions. In experiments involving twelve participants simulating road-crossing decisions under varying traffic volume, marked crosswalks, and traffic signals, neurophysiological analyses reveal hemispheric asymmetries in functional specialization and identify high-beta oscillations (16-25 Hz) in the right fronto-central region (F4) as robust neural markers of motor readiness and decision commitment. Using sliding-window feature extraction, we benchmarked sixteen classification models across traditional, recurrent, and convolutional deep learning architectures, and found that our approach achieved the highest Area Under the Curve (AUC) of 0.727 with an approximate 1-second look-ahead. This work demonstrates how biologically grounded temporal architectures can enhance anticipatory intelligence in autonomous systems and represents the first step toward proactive and adaptive human-robot interaction in the built environment.